Sorbonne Université, CNRS, Laboratoire de Chimie Physique Matière et Rayonnement, UMR 7614, F-75005 Paris, France.
J Chem Phys. 2020 Jun 21;152(23):234103. doi: 10.1063/5.0009264.
Potential energy surfaces (PESs) play a central role in our understanding of chemical reactions. Despite the impressive development of efficient electronic structure methods and codes, such computations still remain a difficult task for the majority of relevant systems. In this context, artificial neural networks (NNs) are promising candidates to construct the PES for a wide range of systems. However, the choice of suitable molecular descriptors remains a bottleneck for these algorithms. In this work, we show that a principal component analysis (PCA) is a powerful tool to prepare an optimal set of descriptors and to build an efficient NN: this protocol leads to a substantial improvement of the NNs in learning and predicting a PES. Furthermore, the PCA provides a means to reduce the size of the input space (i.e., number of descriptors) without losing accuracy. As an example, we applied this novel approach to the computation of the high-dimensional PES describing the keto-enol tautomerism reaction occurring in the acetone molecule.
势能面（PESs）在我们对化学反应的理解中起着核心作用。尽管高效的电子结构方法和代码有了令人印象深刻的发展，但对于大多数相关系统来说，这些计算仍然是一项艰巨的任务。在这种情况下，人工神经网络（NNs）是构建广泛系统的 PES 的有前途的候选者。然而，选择合适的分子描述符仍然是这些算法的瓶颈。在这项工作中，我们表明主成分分析（PCA）是准备最佳描述符集和构建高效神经网络的有力工具：该协议导致神经网络在学习和预测 PES 方面有了实质性的改进。此外，PCA 提供了一种在不降低准确性的情况下减小输入空间（即描述符数量）大小的方法。例如，我们将这种新方法应用于计算描述丙酮分子中酮-烯醇互变异构反应的高维 PES。
